peal.generators.tabular_ddpm

Minimal DDPM for low-dimensional tabular data with guided counterfactuals.

PEAL’s image generators (DiffusionAutoencoder, PathLDM, Stable Diffusion) do not apply to the toy tabular tasks (2-D “circle”-style datasets from peal.data.tabular_datasets). This module provides an MLP score network trained with the standard epsilon-prediction loss, plus an edit method in the EditCapableGenerator protocol that produces counterfactuals by classifier-guided reverse diffusion from a partially noised input. Training writes diffusion.pt, config.yaml and TensorBoard scatter plots into config.base_path.

Classes

BasicDiscreteTimeModel	Score network plus timestep embedding: the trainable part of the DDPM.
PositionalEncoding	Fixed sinusoidal timestep embeddings looked up by integer timestep.
ScoreNetwork	Four-layer MLP predicting the noise eps from x_t and a time embedding.
TabularDDPM	DDPM generator for tabular data with classifier-guided counterfactual edits.
TabularDDPMConfig	Config of TabularDDPM.